Research On Galaxy Image Reconstruction Technology Based On Deep Learnin

The Very Large Array(VLA)is one of the most widely used radio telescopes in the world.In recent years,it has made significant achievements in key astronomical fields such as black hole imaging and galaxy formation,providing important support for human exploration of the universe.VLA uses interferometry and synthetic aperture imaging technology for observation imaging,so the limited number of antennas will lead to insufficient sampling frequency of galactic radiation sources,resulting in point spread function effect,which has a great impact on galaxy imaging.In addition,weak noise from the sky and electronic devices also interferes with the image of the galaxy.At the same time,huge amounts of astronomical data need to be used more quickly and efficiently.For the point spread function effect and noise elimination of radio interference arrays such as VLA,researchers have done a lot of related work,and put forward the Clean reconstruction algorithm and related variants.However,based on the traditional Clean algorithm,the reconstruction effect of the interior texture and the overall contour of the galaxy is not satisfactory in eliminating the point spread function effect containing the extended source galaxy.In this thesis,deep learning algorithm is used to solve the above problems.Firstly,the data sets of different observation parameters under A configuration and D configuration of VLA are constructed,and inspired by relevant deep learning methods and ideas,a MIGAN network based on deep learning is proposed,and its backbone network is a generative adversarial network.By combining the convergent interaction module and the self-interaction module in the generator module,MIGAN focuses on the multi-scale structure and weak information of galaxies,and obtains abstract feature information from different scales more effectively,so as to better eliminate the degradation effect in galaxies with different structures.In the discriminator module,Ge LU is used as an activation function to improve the probability of neuronal output and increase the fitting ability of the network.In the optimization of network parameters,the loss function with constraints,L1 regular loss and perception loss are used as the loss function of MIGAN to update the network weight information,so as to effectively recover the internal detail texture and external contour structure of the galaxy.Through the experiment of galaxy map reconstruction based on deep learning,it is found that the MIGAN network proposed in this paper has a competitive effect on the point expansion function effect elimination of VLA,and has a certain denoising ability.The related image reconstruction quality evaluation index is also better than the traditional Clean algorithm and the current related deep learning network Galaxy GAN.It also proves the applicability of MIGAN network to the point spread function effect and noise cancellation in galaxy maps and the high efficiency of massive astronomical data processing and reconstruction.This method will provide a methodological reference for the elimination of point spread function effect in subsequent astronomical imaging software and radio astronomical arrays.